Running Head: LATENT PROFILE AND LATENT CLASS ANALYSIS

A PRIMER TO LATENT PROFILE AND LATENT CLASS ANALYSIS

Johannes Bauer

University of Erfurt

University of Erfurt

Nordhäuser Str. 63

99089 Erfurt

Germany,

Phone: +49 361 737-2025

Fax: +49 361 737-2019

E-mail: [email protected]

ORCID: 0000-0001-6801-2540

This is a preprint of the following chapter: Johannes Bauer, “A Primer to Latent Profile and
Latent Class Analysis”, published in “Methods for Researching Professional Learning and
Development: Challenges, Applications and Empirical Illustrations”, edited by Michael
Goller, Eva Kyndt, Susanna Paloniemi, and Crina Damşa, 2022, Springer, reproduced with
permission of Springer. The final authenticated version is available online at:
https://link.springer.com/book/10.1007/978-3-031-08518-5
 2

Abstract

This chapter gives an applied introduction to latent profile and latent class analysis

(LPA/LCA). LPA/LCA are model-based methods for clustering individuals in unobserved

groups. Their primary goals are probing whether and, if so, how many latent classes can be

identified in the data, and to estimate the proportional size and response profiles of these classes

in the population. Moreover, latent class membership can serve as predictor or outcome for

external variables. Substantively, LPA/LCA adopt a person-centered approach that is useful for

analyzing individual differences in prerequisites, processes, or outcomes of learning. The

chapter provides a conceptual overview of LPA/LCA, a nuts-and-bolts discussion of the steps

and decisions involved in their application, and illustrative examples using freely available data

and the R statistical environment.

Keywords: Latent variable models, mixture models, latent class analysis, latent profile

analysis.
 3

1 Introduction

Suppose you are a lecturer in a study program who is teaching a compulsory lecture on a

difficult topic. The exam consists of a standardized test that is well-established in your

department and that provides scores ranging between 0 and 50. To get an impression of your

students’ achievement in the recent years you analyze the available data (N = 1000). You are

surprised to find that the data do not follow a normal distribution, as you would have expected,

but a bi-modal one, as illustrated in Figure 1a. Since you are wondering whether something is

wrong with the test you discuss this finding with a colleague. She mentions that the distribution

looks like the data came from two separate groups with different achievement levels. Indeed,

reportedly some students try passing the exam by rote rehearsal whereas others prepare more

thoroughly applying deeper learning strategies. Given the test tasks’ nature, it is reasonable that

rote rehearsal leads to diminished performance. Unfortunately, you do not have data on

students’ learning strategies. All you have is the achievement data. Nevertheless, you wonder

whether it would be possible to identify the unobserved groups given the data.

Let us put the latter question aside for a moment and suppose now we have insight of the

true population characteristics, as illustrated in Figure 1b. Indeed, there are two subpopulations,

the low-achieving students (supposedly superficial learners) marked red and the high achieving

ones (supposedly deep learners) marked blue. Because these subpopulations are unobserved in

our empirical data they are hidden from us and called latent classes. Within each latent class,

achievement differences follow a normal distribution with class-specific (conditional) means

and standard deviations. The bi-modal cross-class distribution is a marginal combination of

these two within-class distributions and, therefore, called a mixture distribution.

 4

Fig. 1. Illustrative univariate finite normal mixture distribution: (a) empirically observed

distribution (N = 1000); (b) true population distributions: two within-class distributions [red:

𝒩 𝜇 15; 𝜎 3 ; blue: 𝒩 𝜇 30; 𝜎 5 ] and their joint mixture distribution (black line).

So, how can we recover the latent classes shown in Figure 1b based on the observed

achievement data? This is where mixture models come in. Mixture models are a family of

statistical procedures for uncovering hidden groups. As Oberski (2016) put it, “Mixture

modeling is the art of unscrambling eggs” (p. 275). In this chapter, I will introduce latent profile

analysis (LPA) and latent class analysis (LCA) which are mixture models for cross-sectional

data. The main difference between them is that LPA applies to continuous response variables,

whereas LCA applies to categorical ones. As we will see, a LPA for the example data in Figure

1a gives us information on (i) the number of latent classes, (ii) model fit and the accuracy of

the classification, (iii) the proportional size of each class, (iv) the within-class means and

variances, and (v) probability values for each person’s latent class membership. Of course, since

we have not observed class membership, the classification is inevitably uncertain. A major

advantage of LPA/LCA over traditional cluster analysis is that this classification uncertainty is

an explicit part of the statistical model (Rost, 2003; Vermunt & Magidson, 2002).
 5

The following section will give readers a conceptual background of LPA/LCA. Next, I

provide didactic examples how to implement LPA and LCA. The chapter will close with some

cautionary considerations, a note on software, and recommendations for further reading.

2 Conceptual overview

2.1 Family relations

As mentioned, LPA/LCA belong to the broad family of mixture models (Hancock &

Samuelsen, 2008; Mclachlan & Peel, 2000; Masyn, 2013; Sterba, 2013). The common purpose

of all mixture models is to provide a probabilistic classification of persons to latent classes

based on a statistical model, where each class is one of the K categories of a discrete latent

variable c. Hence, unlike in traditional factor analysis or IRT modeling, the latent variable is

categorical instead of continuous. All mixture models combine a within-class and a between-

class model (Sterba, 2013). The within-class model defines the population parameters for

persons in a given class k who constitute a relatively homogeneous sub-group regarding these

parameters. The between-class model defines each persons’ probability of being sampled from

the sub-population underlying class k. Thus, mixture models cluster persons who are similar in

terms of the within-class parameters and yield per person a classification probability for each

class. Persons then can be assigned to a latent class based on their highest membership

probability (modal class assignment). This probability-based (soft) clustering is a major

conceptual advantage of mixture models over the deterministic (hard) assignment applied in

traditional cluster analysis because it explicitly accounts for the mentioned classification

uncertainty (Mair, 2018; Rost, 2003; Vermunt & Magidson, 2002).

Differences among various kinds of mixture models concern the parameters that are in focus

of the classification (i.e., conditional means/item probabilities vs. regression parameters) and

whether the data are cross-sectional vs. longitudinal (Tab. 1). Moreover, the observed items’

level of scale can be considered, e.g., to distinguish LPA and LCA. The latter distinction is

somewhat artificial, however, since both LPA and LCA are just special cases of the general
 6

mixture model. It is even possible to combine indicators of mixed levels of scale in the same

analysis (Vermunt & Magidson, 2002; see Mair, 2018, for an illustrative application).

Table 1. Overview of selected mixture models.

Classification by

Means (continuous data) /

Design item probabilities (categorical data) Regression parameters

Cross-sectional Latent profile analysis Regression mixture models

Latent class analysis

Longitudinal Repeated measures latent class analysis Growth mixture models

Latent transition analysis

As indicated in Table 1, latent classes in LPA/LCA are characterized by qualitatively

different response profiles on the analyzed items (Rost, 2003; Vermunt & Magidson, 2002). If

the items are continuous, the profiles are patterns of class-specific means; if they are categorical,

the profiles are defined by the items’ class specific endorsement probabilities (i.e., for

dichotomous items the probability of scoring 1 vs. 0). This basic idea extends to the longitudinal

case. For example, repeated measures LCA models classes of mean/probability-trajectories

across multiple measurement occasions; latent transition analysis models latent classes for each

time point (latent statuses) as well as transitions between them (Collins & Lanza, 2010).

Though not elaborated in this chapter, mixture models for regression coefficients are

noteworthy because they allow uncovering differential treatment effects in experimental or

intervention research (Jung & Wickrama, 2008; Oberski, 2016).

As a final note, mixture models can be used both in an exploratory, hypothesis-generating

mode of data analysis and in a confirmatory, hypothesis-testing one (Finch & Bronk, 2011).

The former occurs more frequently in practice and applies when a researcher has only tentative
 7

or no prior assumptions about the latent classes. It should be emphasized, however, that even

confirmatory LPA/LCA cannot test hypotheses about a specific number of classes. As

elaborated below, this number is not an estimable model parameter. Testable hypotheses

concern only the class sizes and their response patterns on the items.

2.2 The LPA and LCA models

LPA. In LPA, we typically assume a normal distribution within each latent class, though

other distributions are possible, such as Poisson for count data (Mair, 2018; Vermunt &

Magidson, 2002). Note that the normality assumption applies only to these subpopulation

distributions, not to their joint distribution, and it is untestable. Each latent class (1, …, K) has

a density 𝑓 𝑥 |𝜇 , 𝜎 , where 𝑥 is an observed value for person i on a continuous indicator m,

and 𝜇 and 𝜎 are its mean and variance in class k. The joint mixture model is given by

𝑓 𝑥 𝜋 𝑓 𝑥 |𝜇 , 𝜎 (1)

where 𝜋 is the relative class size, or equivalently, the class probability (also called mixture

weight). That is, each person’s density in the mixture distribution is a sum of the class specific

normal densities weighted by the class probabilities. For example, consider persons with a test-

score of 21 in Figure 1b. Their mixture density (i.e., value on the black curve) is the sum of the

two class-specific densities (i.e., values on the red and blue curves, respectively) which are

weighted by the respective class probabilities. In the multivariate case, where 𝒙𝒊 is a vector of

responses on a set of M observed items, the model remains essentially the same but each class

has its specific mean vector 𝝁𝒌 and variance-covariance matrix 𝚺𝒌 .

Of course, before the analysis, we neither know the class probabilities nor the class-specific

means and (co)variances. Unless further restrictions are imposed, these are the model

parameters that have to be estimated. Because the number of free parameters grows quickly

with the number of latent classes and indicators, it makes sense to specify further restrictions to

increase model parsimony and estimation stability (Vermunt & Magidson, 2002). Frequently,
 8

one imposes a local independence restriction meaning that the indicators are constrained to be

uncorrelated within each latent class.1 Additionally, a cross-class equality restriction on the

indicator variances can be imposed delivering homogeneity of the variance-covariance matrices

across classes (Vermunt & Magidson, 2002). Together these two restrictions ensure that the

classes differ only in their mean profiles, and not in the form of the variance-covariance matrix,

which facilitates interpretation (Peugh & Fan, 2013). Several other models with different sets

of restrictions are available (Masyn, 2013; Pastor et al., 2007; Scrucca et al., 2016). Finally, an

important point visible in equation 1 is that the number of classes is not an estimated model

parameter. Rather, the number of classes has to be decided upon by the analyst on grounds of

statistical and substantive criteria (see section 2.3). Once a target model has been selected, its

estimated parameters can be interpreted.

LCA. The LCA model works similarly though it applies to categorical items. LCA starts

with the basic idea of local independence, that is, it seeks to establish latent classes such as that

responses on a set of categorical items are independent within classes (Rost, 2003). Suppose

we have a vector of responses 𝒖𝒊 to a set of M categorical items. Under the local independence

assumption, the LCA model provides the probability for a given response pattern on these items

by

𝑃 𝒖𝒊 𝜋 𝑃 𝑢 |𝑐 𝑘 . (2)

That is, the probability of a specific combination of item responses depends on the class

probabilities and the product of the item response probabilities given class membership. These

are the model parameters to be estimated. For items with J categories the item probabilities are

the J-1 thresholds between these categories. Again, further restrictions may be imposed on the

1
That is, residual correlations among the indicators are zero given class membership. Local independence is
a default assumption in many latent variable models, such as factor analysis or IRT models, but can be relaxed.
 9

model. Note that the local independence assumption can be evaluated and also be partially

relaxed in LCA (Asparouhov & Muthén, 2015; Masyn, 2013).

2.3 Steps in LPA/LCA

Conducting an LPA/LCA is an iterative, stepwise process similar to exploratory factor

analysis. Four steps can be distinguished, the last of which is optional: (i) specifying the model,

(ii) determining the number of latent classes, (iii) interpreting the target solution substantively,

and (iv) including predictors and/or outcomes of latent class membership. In the following, I

briefly elaborate on each of these steps.

Step 1: Model specification. In specifying the model researchers need to make a theory-

based selection of observed indicators, decide how these indicators enter the model, and what,

if any, restrictions to impose on the model parameters. Concerning the selection of indicators,

one should keep in mind that LPA/LCA are measurement models (Wang & Wang, 2020). So,

there should be a rationale that the indicators measure a joint theory-based construct (e.g.

different profiles of self-concept). In practice, LPA/LCA is also frequently applied to classify

persons over a range of conceptually related, but distinct constructs (e.g., J. Bauer & Mulder,

2013; Gillet et al., 2020). In that case, the resulting profiles are mere summaries of persons with

similar answer patterns and should not be interpreted as realizations of a latent trait (see section

4.1 on the danger of reification). In either case, it makes only sense to apply LPA/LCA if the

set of indicators is inter-correlated. So, screening item associations prior to the analysis is

warranted.

A further consideration in LPA is whether to include indicators on the item level or in

aggregated form, and whether to apply prior transformations. Aggregation is frequently done

by computing composite mean scores, factor scores, or (preferably) plausible values of the

(sub-)scales to be analyzed (e.g., Bauer & Prenzel, 2021; Hong et al., 2020; Kunst et al., 2016;
 10

Mair, 2018).2 This approach considerably reduces the number of parameters to be estimated

and better approximates continuous indicators. An arguably optimal approach in this regard is

second-order LPA in which the latent class variable is a second-order variable with continuous

first-order factors serving as indicators (for an application see Bauer et al., 2018). With this

specification, the indicators used for the latent classification are latent factors that are truly

continuous and corrected for measurement error. Note that this approach requires scaling the

first-order factors to have a meaningful mean structure (Little et al., 2006).

Concerning parameter restrictions, researchers may wish to start with the local

independence model since this is the traditional basis of LCA, and potentially consider cross-

class equality constraints on the indicator variances. Any other restrictions should be made on

substantive grounds (see Rost, 2006, for interesting applications). Attention should be payed to

the default restrictions implemented by specific software packages. For example, Mplus

(Muthén & Muthén, 1998-2017) by default imposes local independence and homogeneity

across classes. An alternative strategy implemented routinely in the mclust package (Scrucca

et al., 2016) is estimating a variety of models with different restrictions and choosing the best

fitting one (including a decision on the number of latent classes; see Step 2). This data driven

strategy alleviates researchers from a priori considerations about restrictions. However,

interpreting what the restrictions applied by the selected model mean substantively may not

always be straightforward. For further details, see section 3.2.

Step 2: Class enumeration. Deciding on the number of classes may be the most difficult

step in the whole analysis. The goal is to yield a solution that balances model parsimony and

fit, and that delivers substantively interpretable classes. For this purpose, one estimates3 a series

2
Using composite scores of categorical items as indicators will turn a LCA into a LPA.
3
Discussing details of estimation is beyond the scope of this chapter; for a tractable introduction, see Masyn
(2013).
 11

of models from 1 to K latent classes and compares them on indices of relative model fit and

classification diagnostics. Then, the candidate model(s) are considered substantively.

The most commonly used fit indices are the Bayesian Information Criterion (BIC), its

sample-size-adjusted variant (SABIC), and the Akaike Information Criterion (AIC). With these

and any other information criteria, lower values indicate better model fit. Ideally, a minimum

value in the series of candidate models points at the best-fitting solution. In practice, fit indices

may continue to improve as more latent classes are extracted. In such circumstances, an elbow-

plot – similar to the scree-plot in exploratory factor analysis – can be useful to identify at which

step the improvement in model fit plateaus. Next to information criteria, models with adjacent

numbers of latent classes can be compared by a likelihood ratio test (LRT) which is available

in several versions. Some simulation studies indicate the bootstrapped LRT (BLRT) to be quite

effective in class enumeration (Nylund et al., 2007). A statistically significant result of the LRT

supports the k-class model over k-1 classes. In theory, one choses the model after which the

LRT is statistically non-significant for the first time. Again, in practice the LRT may continue

to stay significant for the whole series of probed models, especially with large sample sizes.

Classification diagnostics can additionally inform the class enumeration process. For each

latent class, average posterior class probabilities (AvePPk) provide a measure of classification

accuracy. As the name suggests, they are based on averaging for class k the model-estimated

(i.e. posterior) latent class membership probabilities for individuals with their highest

membership probability in this class. Values ≥ .70 have been proposed as desirable (Masyn,

2013). The information about classification accuracy is condensed in the entropy measure that

ranges between 0 and 1 and can be interpreted similar to a reliability coefficient for the

classification (values ≥ .80 are desirable, ≥ .60 a suggested minimum; Asparouhov & Muthén,

2014). Classification accuracy may be helpful to decide between solutions that have similar fit

(Rost, 2006), but note that entropy is not a good sole or primary model selection criterion

(Masyn, 2013).
 12

As a final word of caution, though there is extensive research on fit indices in mixture

models (e.g., Nylund et al., 2007; Peugh & Fan, 2013; Tofighi & Enders, 2008), results have

not yet converged and there is no single best index. Applied researchers frequently combine an

array of indices for class enumeration but – probably more often than not – their results may

prove to be inconsistent. Therefore, there is general agreement in the methodological literature

that model selection needs to be guided at least as much by the substantive interpretability of

the class solution as by fit indices. If multiple solutions with similar fit occur, one should inspect

all of them and make a decision on substantive grounds.

Step 3. Substantive interpretation of the target model(s). The interpretation process in

LPA/LCA is similar to traditional cluster analysis or exploratory factor analysis. For this

purpose, one inspects the class-specific mean/probability profiles across the indicators and

differences in these profiles across classes. Depending on model specification, class specific

(co)variances are of importance, too. Based on this interpretation, a definition and label need to

be attached to each class. Finally, the class sizes are of interest to distinguish common or

normative profiles from more exceptional ones. In general, no solution should be retained that

is not well interpretable, regardless of model fit. Some authors consider a solution useless in

which not all classes are interpretable (Wang & Wang, 2020). However, Mair (2018) notes that

the occurrence of a single uninterpretable junk-class is not problematic since it may simply

contain the persons that did not fit in any of the other classes.

Though interpretation is primarily a theoretical issue, some methodological considerations

apply (Collins & Lanza, 2010; Masyn, 2013; Rost, 2006): First, the class profiles should be

distinct enough. That is, there should be low variability within classes and a high one between

them (Collins & Lanza, 2010). In this context, it may be helpful to compare solutions with

adjacent numbers of classes to see from what classes in a prior solution members of a new class

are drawn (Rost, 2006). Sometimes a class in the prior solution is simply split up in two classes

and the researcher needs to consider the value added by this differentiation. Moreover, finding
 13

the structure of profiles to vary greatly across different solutions may be indicative that

LPA/LCA is not an appropriate model for the data (Rost, 2006).

Second, it should be checked, whether the indicators can be arranged in a way so that the

class profiles are ordered (i.e., they do not cross each other in a profile plot). Such a solution –

particularly, if the profiles are more or less parallel – would indicate that the classes differ by

degree, rather than by type (Rost, 2006; Wang & Wang, 2020). In such a case, the classification

is still meaningful, but a model focusing on quantitative differences in the latent variable (i.e.,

a factor or IRT model) would provide a more parsimonious account of the data.

Finally, solutions with several small classes (say ≤ 5%) may be indicative that too many

classes have been extracted (Nylund et al., 2007). Though such small classes can be of

substantive interest, it should be checked critically whether they have a distinct and

interpretable profile.

Step 4: Include predictors and distal outcomes of most likely latent class membership.

Having established a latent class solution, a logical next step is to analyze predictors or

dependent variables of latent class membership (Vermunt & Magidson, 2002). This is called

LPA/LCA with covariates and distal outcomes. Such analyses are useful (i) to enrich the

description of the latent classes based on background information (e.g., social background

variables or experimental condition), (ii) to analyze the construct-validity of the latent class

solutions, or (iii) to test hypotheses for answering substantive research questions. The latter two

cases lead to a blend of exploratory and confirmatory analyses (Masyn, 2013) because the

hypotheses can only be stated after the classification. However, even if the classification is

intrinsically exploratory, the later hypotheses about the classes’ relation to covariates and distal

outcomes can still be tested in a confirmatory fashion.

Methods for analyzing covariates and distal outcomes in mixture models are still an active

area of research. Though it may be tempting to simply extract individuals’ modal class

assignment from the results and use this as a manifest variable for further analyses (e.g. as factor
 14

in an ANOVA), it is well known that this procedure leads to biased results and is inappropriate

(Bakk & Kuha, 2020). The reason is that this hard classification disregards the classification

uncertainty. A variety of alternative approaches has been developed. Recent overviews

recommend using the two-step method or one of the bias-corrected three-step methods (BCH

and ML) for handling covariates; (manual) BCH is advisable for continuous and categorical

distal outcomes (Asparouhov & Muthén, 2021; Bakk & Kuha, 2020; Nylund-Gibson et al.,

2019). Unfortunately, these approaches are currently only available in commercial software

(Latent Gold, Mplus). Illustrating the application of LPA/LCA with covariates and distal

outcomes is beyond the scope of this chapter, but readers will find didactic examples in the

literature cited above as well as in Wang and Wang (2020).

2.4 A person-centered conceptual perspective

So far, this chapter has introduced LPA/LCA from a statistical perspective. What, however,

is the substantive added value of mixture models – and LPA/LCA in particular – for studying

professional learning and development? Mixture models have unique advantages in studying

individual differences in learning, its conditions, and outcomes (Bauer et al., 2018; Collins &

Lanza, 2010; Hancock & Samuelsen, 2008; Masyn, 2013). Conceptually, mixture analyses

adopt a person-centered approach that is helpful when a researcher aims to analyze categorical

individual differences in terms of a classification of persons according to typical patterns of

their characteristics or behavior (Bergman & Magnusson, 1997). This person-centered

perspective is complementary to the traditional variable-centered one that aims at identifying

relationships between variables. For example, taking a variable-centered perspective,

researchers might be interested in the degree to which some predictors are relevant for

explaining employees’ engagement in work-related learning activities. From a person-centered

perspective they may ask what typical patterns of engagement in learning activities occur, how

prevalent these patterns are, and what background variables characterize people demonstrating

a specific pattern. Hence the two perspectives are complementary and should not be
 15

dichotomized falsely (Masyn, 2013). Synergies can arise from taking both perspectives on the

same research topic (Marsh et al., 2009). For example, Bauer and Mulder (2013) studied

predictors of and individual differences in nurses’ engagement in learning activities after errors

at work. In a variable-centered analysis, they found that nurses’ perceptions of errors and their

social context predicted this engagement. A subsequent LPA complemented these findings by

identifying four typical patterns of reactions to errors and learning engagement, with a

substantial proportion of nurses (> 40%) showing dysfunctional behavior.

More generally speaking, LPA/LCA can be used whenever there are theoretical reasons to

believe that a latent construct is categorical. As Collins and Lanza (2010) argue, many variables

of substantive interest can be understood both in terms of continua and categories. They employ

the example of alcohol consumption which may be defined continuously as the volume of

alcohol consumed in a given period of time. At the same time, it makes substantive sense to

distinguish categories of alcohol consumption, such as non-drinking, occasional drinking,

social drinking, binge drinking, or being alcohol addict. Referring to our introductory example,

test achievement is arguably a continuous variable, but it can be meaningful to distinguish

learners showing mastery vs. non-mastery. So, readers will find that LPA/LCA (and other

mixture models) are applicable to a broad spectrum of variables relevant to professional and

work-related learning.

3 Implementation of LPA and LCA

This section will provide three didactic examples of LPA and LCA. We will use simulated

data and the packages mclust (Scrucca et al., 2016), tidyLPA (Rosenberg et al., 2018), and

poLCA (Linzer & Lewis, 2011) in the R statistical environment (R Core Team, 2021) so that

readers can reproduce all analyses. Though the data are simulated, Examples 2 and 3 are based

on published studies that used LPA (Gillet et al., 2020) and LCA (Richter et al., 2013). Below,

I restrict the presentation to the most important parts of the in- and output. Complete code and

data files are available in the supplemental online material (SOM) at https://osf.io/x3ahn.
 16

3.1 Example 1: Univariate LPA of student achievement

Turning back to the initial example from Figure 1, we will use LPA to recover the two

hidden groups in the mixture distribution of student achievement. Since there is only one

variable, our only choice in model specification is whether we allow the class-specific variances

to vary, or not. Since our empirical data suggest different spread in the presumed clusters (Fig.

1a) it seems reasonable to freely estimate the variances within each class.

For class enumeration, we estimate models with k = 1 to 4 latent classes using the code

pasted below. The argument modelName = "V" specifies varying variances across classes.

If we omit this argument, mclust will estimate models with and without equality constraints

and users can compare their fit to the data. The next commands call for an elbow plot of BIC

(see SOM) and the BLRT.

library(mclust)
m1 <‐ Mclust(data = mix, modelName = "V", G = 1:4)
plot(m1, what = "BIC")
LRT <‐ mclustBootstrapLRT(mix, modelName = "V", maxG = 3)

Unfortunately, mclust delivers only a few fit indices. If we desire a broader variety of

indices, we can compute them manually based on the information stored in the output object

m1. Otherwise, we may get enhanced output by using the functions

estimate_profiles() and get_fit() from tidyLPA which provides a set of wrapper

functions for mclust. Both options are demonstrated in the SOM. Table 2 presents a selection

of fit indices, all of which indicate that the two-class model fits the data best. This is backed up

by the BLRT showing that the two-class model fits the data significantly better than the one-

class model, whereas moving to three classes does not provide a further statistically significant

improvement in fit.
 17

Table 2. Selected fit indices for the univariate LPA of student achievement.

Classes LL AIC BIC CAIC SABIC BLRT p

1 -3519.63 -7043.26 -7053.08 -7055.08 -7046.73 --- ---

2 -3399.82 -6809.64 -6834.18 -6839.18 -6818.30 239.63 .001

3 -3400.23 -6816.46 -6855.72 -6863.72 -6830.31 -0.82 .946

4 -3398.41 -6872.80 -6837.86 -6883.80 -6818.82 3.65 .135

Note. LL = Log Likelihood; note that in mclust higher values on information criteria indicate

better fit; boldface = highest value/first non-significant BLRT.

Concerning classification diagnostics, we can use the uncertainty values and modal class

assignments stored in the mclust output object to calculate average posterior class probabilities

which are AvePP1 = .95 for Class 1 and AvePP2 = .98 for Class 2 (see SOM). Entropy values

can be obtained by get_fit()and show that the 2-class model has high classification

accuracy (E = .89). Thus having identified a best fitting model, we can inspect its estimated

parameters in the output pasted below.

summary(m1, parameters = TRUE)

## ‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
## Gaussian finite mixture model fitted by EM algorithm
## ‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
##
## Mclust V (univariate, unequal variance) model with 2 components:
##
## log‐likelihood n df BIC ICL
## ‐3399.819 1000 5 ‐6834.176 ‐6903.184
##
## Clustering table:
## 1 2
## 309 691
##
## Mixing probabilities:
## 1 2
## 0.3090278 0.6909722
##
## Means:
## 1 2
## 15.23148 30.14721
##
 18

## Variances:
## 1 2
## 9.633501 23.584189

The first part of the output reiterates information about model fit. The clustering table tells

us that n = 309 persons had their modal class membership in Class 1, and n = 691 in Class 2.

The estimated class sizes in the population (mixing probabilities) are 30.9 % for Class 1 and

69.1 % for Class 2. The most important part of the output are the sections on class-specific

means and variances. Class 1 has an average achievement of M = 15.2 (SD = 3.1), and Class 2

M = 30.1 (SD = 4.9). Given these values, the two classes can be interpreted as consisting of

low-achievers and high-achievers, respectively.

Apparently, our LPA recovered the class-specific population parameters very well (cf. Fig

1b). Of course, LPA cannot deliver evidence that there are two latent classes because some

students applied different learning strategies than others. To investigate this, we might conduct

a new study collecting additional data on students’ learning strategies and checking whether

strategies predict latent class membership. We could also collect additional data on some

external criterion (say later study success) and use the latent class variable as a predictor.

Finally, with our simulated data, it is possible to compare class assignment based on the

LPA results with individuals’ true class membership. As table 3 shows, there is a correct

classification rate of 97.5%. So, the LPA was quite effective in recovering the unobserved

groups.

Table 3. True class versus most likely class membership from LPA.

LPA

True 1 2 Total

A 292 8 300

B 17 683 700

Total 309 691 1000

 19

This example of a univariate LPA demonstrated general procedures that transfer to the

multivariate case. Though univariate LPA may rarely be of substantive interest in applied

research, it can be a useful methodological tool. For example, in a recent online study, we used

it to identify a cluster of participants with unreasonably short time on task that we discarded as

junk participations. LPA provides arguably a more principled approach for this purpose than

setting an arbitrary cut off time. Moreover, univariate LPA may be an alternative to common

grouping procedures, such as the median split (Rost, 2006). For this purpose, equality

restrictions may be imposed on the class sizes to yield equally sized groups.

3.2 Example 2: Multivariate LPA of work recovery

In most applied cases researchers will wish to use LPA on multiple indicators. The

illustrative example presented in this section is based upon Study 1 in Gillet et al. (2020) who

investigated profiles of work recovery (and related predictors) in a sample of N = 415

employees from various occupations. Factor scores on three psychological mechanisms of work

recovery served as indicators: (i) psychological detachment as the ability to mentally step back

from work after hours; (ii) rumination as a tendency to repetitively or continuously be

preoccupied with distressful aspects after work; and (iii) overcommitment as a generic tendency

to excessively strive for achievement and recognition at work. Whereas psychological

detachment is a functional mechanism for the work recovery process, rumination and

overcommitment are dysfunctional ones. Population values for the data simulation (N = 500)

were chosen in line with results from Gillet et al. (2020). Hence, within sampling error, our

analysis will lead to roughly identical conclusions.

The code below instructs mclust to specify and fit a series of models with k = 1 to 8 latent

classes. As mentioned above, if we specify no particular model restrictions the software

estimates a set of 14 models that impose specific model constraints based on different geometric

features of 𝚺𝒌 (Scrucca et al., 2016). The substantive implications of these features may be hard
 20

to interpret for applied researchers. Important models, also discussed in other literature on LPA

(Masyn, 2013; Pastor et al., 2007; Rosenberg et al., 2018), are in order of decreasing parsimony:

 EEI: variances equal across classes, covariances fixed at zero

 VVI: class-varying variances, covariances fixed at zero

 EEE: variances and covariances equal across classes

 VVV: class-varying variances and covariances

Of these, EEI and VVI are local independence models, EEE and VVV relax this assumption.

m1 <‐ Mclust(wr, G = 1:8)

summary(m1)

## ‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
## Gaussian finite mixture model fitted by EM algorithm
## ‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
##
## Mclust VVI (diagonal, varying volume and shape) model with 4 components:
##
## log‐likelihood n df BIC ICL
## ‐1450.855 500 27 ‐3069.505 ‐3104.645
##
## Clustering table:
## 1 2 3 4
## 58 159 225 58

The summary output suggests that the VVI model with four latent classes fits the data best.

To inspect fit more closely, we can call for an elbow-plot comparing all mclust models (see

SOM). Of course, as an alternative to this data-driven strategy, we could have specified a priori

parameter restrictions on 𝚺𝒌 based on substantive considerations. For example, we could have

specified local independence and equal cross-class variances by setting the argument

modelName = "EEI". To test the tenability of these restrictions, next, we could compare

the fit of the EEI model to a less restrictive one such as VVI.

As demonstrated in the univariate example, it may be desirable to base the class

enumeration decision on a broader set of indices. Again, this is achieved conveniently using

tidyLPA. The function compare_solutions() from this package delivers a combined

 21

recommendation based on a number of fit indices (see SOM). Moreover, it may be helpful to

plot selected indices. As visible from Figure 2, the four-class solution fits best according to

BIC, CAIC. AIC and SABIC seem less conclusive but parsimony would speak for the four-

class solution, too, unless a model with more classes would have advantages in terms of

substantive interpretability. The BLRT pasted in the output below adds further support to the

four-class solution by getting non-significant for the comparison of 5 versus 4 classes.

Fig. 2. Elbow-plot of selected fit indices for the multivariate LPA.

set.seed(123456)
LRT <‐ mclustBootstrapLRT(wr, modelName = "VVI", maxG = 5)
LRT

## ‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
## Bootstrap sequential LRT for the number of mixture components
## ‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
## Model = VVI
## Replications = 999
## LRTS bootstrap p‐value
## 1 vs 2 551.686999 0.001
## 2 vs 3 289.270784 0.001
## 3 vs 4 245.745002 0.001
## 4 vs 5 18.770582 0.051
## 5 vs 6 3.532327 0.937
 22

Next, we inspect classification diagnostics. The listing below delivers AvePP values all of

which are satisfactory. Entropy is high, too (E = .93; see SOM). mclust offers additional

diagnostic plots based on dimensionality reduction that allow inspecting the boundaries

between the classes to see whether they are well separated (SOM).

round(aggregate(x = 1 ‐ m1$uncertainty,
by = list(m1$classification),
FUN = "mean"), 2)

## Group.1 x
## 1 1 0.96
## 2 2 0.96
## 3 3 0.97
## 4 4 0.98

Having decided on the number of classes, we can refit the target model and inspect the

model results. Below is a shortened output.

m1.4 <‐ Mclust(wr, modelName = "VVI", G = 4)

summary(m1.4, parameters = TRUE)

(...)

## Clustering table:
## 1 2 3 4
## 58 159 225 58
##
## Mixing probabilities:
## 1 2 3 4
## 0.1138546 0.3152577 0.4546402 0.1162474
##
## Means:
## [,1] [,2] [,3] [,4]
## overcom ‐1.540950 0.5203160 ‐0.4246870 1.636629
## pdetach 1.417878 ‐0.4843335 0.3818300 ‐1.548249
## rumin ‐1.264305 0.4850927 ‐0.3714648 1.185387
##
## Variances:
## [,,1]
## overcom pdetach rumin
## overcom 0.1200567 0.00000000 0.0000000
## pdetach 0.0000000 0.05935828 0.0000000
## rumin 0.0000000 0.00000000 0.4626966
## [,,2]
## overcom pdetach rumin
## overcom 0.1068036 0.0000000 0.0000000
 23

## pdetach 0.0000000 0.1490478 0.0000000

## rumin 0.0000000 0.0000000 0.1430751

(...)

The mixing probabilities indicate that class sizes vary between 11 and 45 % of the sample.

The class-specific mean profiles can be found under “Means”. Plotting them facilitates their

interpretation (Fig. 3). As concluded in Gillet et al. (2020), the classes can be interpreted to

demonstrate high (Class 1, 11 %), moderately high (Class 3, 45 %), moderately low (Class 2,

32 %), and low (Class 4, 11 %) ability to achieve work recovery. Note that this implies that the

identified latent classes differ mostly in terms of degree rather than by type. Indeed, the profiles

would be perfectly ordered had we recoded psychological detachment so that higher values on

all three indicators meant worse ability to achieve work recovery. As discussed above, such an

ordered classification is meaningful by pointing at clusters of persons who show similar levels

of the investigated trait(s). Nevertheless, these quantitative differences could be modeled more

parsimoniously using a continuous latent variable model (Rost, 2006; Wang & Wang, 2020).

So, some might question whether LPA is the best modeling choice in the present case.

Fig. 3. Mean profiles of four latent classes of work recovery (simulated data based on Gillet et

al., 2020).
 24

The last part of the output shows the class specific variance-covariance matrices. As implied

by the VVI model, the class-specific covariances are fixed at zero (i.e. local independence) and

the indicator variances are allowed to vary. For example, the variance of rumination is higher

in Class 1 than in Class 2. Such information may be important for understanding the nature of

the latent classes.

Finally, a comparison with individuals’ true population class membership shows that about

97 % of our sample have been classified correctly by the LPA (SOM).

3.3 Example 3: LCA of teachers’ participation in professional development

This section gives an applied example of LCA for categorical variables using the poLCA

package. It draws upon a study by Richter et al. (2013) who investigated individual differences

in language teachers’ (N = 2076) participation in professional development (PD) programs. In

their LCA, the authors analyzed eight dichotomized items indicating whether teachers during

the last five years had participated (yes/no) in PD activities on each of the following topics: (i)

content knowledge, and (ii) pedagogical content knowledge, each with a focus on reading

comprehension, (iii) teaching and learning methods, (iv) teaching heterogeneous students, (v)

classroom management and discipline, (vi) individualized learning support, (vii) counseling

students and parents, and (viii) knowledge about kids and youth. As in Example 2, we will use

simulated data (N = 2076) with population parameters based on the outcomes of the original

study.

For model specification, poLCA requires users to explicitly provide a model formula. The

first part of the code below specifies a LCA-model with no covariates and stores it in an R

object. Next, we use it to estimate a series of models with k = 1 to 7 latent classes. Unlike in

mclust, each model needs to be estimated in a separate function call. The argument nrep

specifies the number of times to estimate each model using different sets of random starting

values for the estimation algorithm. This is important to avoid solutions that converge at local

instead of global maxima of the likelihood function. Local maxima are a plague in mixture
 25

modeling (e.g. Masyn, 2013). We did not need to take care of this problem in the previous LPA

examples because mclust employs a different approach to avoid local maxima (Scrucca &

Raferty, 2015). In the present LCA, we estimate each model with at least ten sets of random

starting values. Larger numbers are typically required for complex models, especially in serious

applications. Here, we increase the number of random starts for the models with 6 and 7 latent

classes and additionally increase the maximum number of iterations (default = 1000) to attain

convergence of the EM algorithm.

library(poLCA)
f <‐ cbind(CK, PCK, TLM, THS, CMD, ILS, CSP, KKY) ~ 1
set.seed(934857)
m1 <‐ poLCA(f, pd[ ,‐9], nclass = 1, nrep = 10)
m2 <‐ poLCA(f, pd[ ,‐9], nclass = 2, nrep = 10)
m3 <‐ poLCA(f, pd[ ,‐9], nclass = 3, nrep = 10)
m4 <‐ poLCA(f, pd[ ,‐9], nclass = 4, nrep = 10)
m5 <‐ poLCA(f, pd[ ,‐9], nclass = 5, nrep = 10)
m6 <‐ poLCA(f, pd[ ,‐9], nclass = 6, nrep = 50, maxiter = 5000)
m7 <‐ poLCA(f, pd[ ,‐9], nclass = 7, nrep = 50, maxiter = 5000)

To assess model fit, poLCA delivers the information criteria AIC and BIC. As in the

previous examples, further fit indices can be calculated manually based on the model results

(SOM). The elbow plot in Figure 4 shows that BIC, CAIC, and SABIC point at a best fitting

solution with five latent classes whereas AIC continues to improve as more classes are

extracted. Unfortunately, the BLRT is not available in poLCA. So, based on the present

information, we may preliminarily decide for the 5-class model and inspect classification

diagnostics.
 26

Fig. 4. Elbow-plot of selected fit indices for the LCA.

The listing below, first calculates AvePP all of which are > .70. Regarding entropy, note

that poLCA does not estimate relative entropy as is typically used in other software. The

function provided below provides this index (cf. Collins & Lanza, 2010, p. 75). As we can

obtain from the output, entropy for the 5-class model is somewhat below the recommended

value of .80 but acceptable (i.e., > .60). From the series of values, we can see that initially

entropy decreases slightly as more classes are extracted. As Collins and Lanza (2010) note, this

is frequently the case and an artefact of chance. Recall from the discussion above that entropy

is not a good indicator for class enumeration.

round(aggregate(x = m5$posterior, by = list(m5$predclass), FUN = "mean"), 2

)

## Group.1 V1 V2 V3 V4 V5
## 1 1 0.75 0.06 0.09 0.09 0.01
## 2 2 0.01 0.88 0.02 0.03 0.06
## 3 3 0.03 0.03 0.89 0.04 0.00
## 4 4 0.06 0.06 0.05 0.78 0.05
## 5 5 0.00 0.13 0.00 0.03 0.84

lca_entropy <‐ function(x) {

1 ‐ ((sum(‐x$posterior*log(x$posterior)))/(nrow(x$posterior)*log(ncol(x$pos
terior))))

}
 27

lca_models <‐ list(m2, m3, m4, m5, m6, m7)

entropy <‐ sapply(lca_models, FUN = lca_entropy)
round(entropy, 2)

## [1] 0.83 0.78 0.76 0.76 NaN 0.79

Having decided for the 5-class model, we can inspect the model parameters. Below is a

shortened output.

m5r

## Conditional item response (column) probabilities,

## by outcome variable, for each class (row)
##
## $CK
## Pr(1) Pr(2)
## class 1: 0.0217 0.9783
## class 2: 0.0355 0.9645
## class 3: 0.6953 0.3047
## class 4: 1.0000 0.0000
## class 5: 0.8745 0.1255
##

(...)

##
## Estimated class population shares
## 0.2908 0.3073 0.143 0.0718 0.1872
##
## Predicted class memberships (by modal posterior prob.)
## 0.3179 0.2813 0.1392 0.0737 0.1879

(...)

In the first part of the output, poLCA gives us the class-specific response probabilities per

indicator and category. For example, teachers in Class 1 have a 98 % probability of having

attended PD focusing on content knowledge, whereas teachers in Class 5 have only a 13 %

probability. Next, the output contains the latent class sizes in two forms: as estimated class

population shares and based on modal class assignment. Both types of values are quite close as

they should be. For class interpretation, poLCA provides a 3D-barplot of the conditional item

response probabilities (SOM). This plot is hard to read, however, especially with growing

numbers of classes, indicators, and answer categories. Since we analyzed dichotomous

 28

variables, a profile plot of the conditional item probabilities is better suited for interpretation

(Fig. 5).

Fig. 5. Conditional item probabilities of five latent classes of participation in professional

development activities (simulated data and adapted figure based on Richter et al., 2013); CK =

content knowledge, PCK = pedagogical content knowledge, TLM = teaching and learning

methods, THS = teaching heterogeneous students, CMD = classroom management and

discipline, ILS = individualized learning support, CSP = counseling students and parents, KKY

= knowledge about kids and youth.

Following Richter et al. (2013) the classes can be interpreted as high on all activities (Class

1, 29.1 %), high on activities focusing (pedagogical) content knowledge and teaching methods

(Class 2, 30.7 %), high on activities focusing general pedagogical topics (Class 3, 14.3 %),

high on activities focusing teaching methods (Class 4, 7.2 %), and low on all activities (Class

5, 18.7 %). Note that, due to sampling error, the class sizes estimated from our simulated data

deviate somewhat from Richter et al. (2013).

 29

Two more points are noteworthy. First, the order of classes in any mixture model is arbitrary

and, thus, may change over a series of models, software used etc. That is, Class 1 from a given

solution, may be Class k in another. The order of classes can be adjusted by providing respective

starting values to the estimation algorithm (SOM). In the above output, classes are reordered to

match results reported in Richter et al. (2013). Second, a closer inspection the output reveals

that Class 4 has boundary conditional probabilities on the indicator CK (1 and 0 for no and yes,

respectively). Finding many such boundary estimates may be a warning sign that the model

solution is invalid or that too many classes have been extracted (Geiser, 2013). Even a single

one should be a cause for caution. In general, boundary values may be legitimate estimates but

they can also be an indication of model misspecification. In the present example, the respective

population value for the simulation was p = .98 which already is close to the upper bound of

probability. So, within sampling error, the result is a reasonable estimate.

Finally, a check of how well LCA recovered individuals’ true class membership shows only

an 85 % correct classification rate with misclassification being particularly apparent for Class

4 (SOM). This reduced accuracy may be due to the classes not being well separated from the

outset, as hinted by the suboptimal entropy value. Hence, this example may serve as a caveat

that even in solutions that seem reasonable in terms of fit and interpretation a non-trivial degree

of misclassification can be prevalent.

4 Conclusion

This chapter provided an applied introduction to LPA/LCA and their implementation in

statistical software using simulated examples of real research. As demonstrated, LPA/LCA are

effective analytic techniques for identifying unobserved subgroups that can help researchers to

investigate individual differences in learning, among other interests. Modern software as well

as methodological advances have greatly facilitated the use of these analytic techniques for

applied researchers.
 30

4.1 Some words of caution

Despite the great potential of mixture models some cautionary words are in order. First, as

elaborated in this chapter, some aspects of mixture modeling are inevitably exploratory. This

should humble researchers in drawing overly strong conclusions from them. In the same vein,

readers will have recognized that mixture modeling is a complex process that requires a

multitude of decisions and, thus, provides great researcher degrees of freedom. The analytic

results can be quite sensitive to these choices and decisions. Therefore, it is crucial to work out

a solid and principled plan ahead of the analyses, including a specification of the criteria for the

various decisions to make. Otherwise, there is a substantial danger of getting lost in the analytic

garden of forking paths.

Second, several fallacies in the interpretation of mixture models are to avoid. As Loken and

Molenar (2007) note, “(…) successfully applying a mixture model does not immediately

confirm the appropriateness of the model” (p. 278). That is, finding a relatively best fitting

model within a series of candidate models neither implies that the model is statistically good in

an absolute sense nor that it is a substantively correct representation of the phenomenon in the

population. Moreover, results from a mixture model cannot provide sufficient evidence to

confirm that the construct of interest indeed is categorical in nature (Loken & Molenar, 2007;

Masyn, 2013). Above I have emphasized that, conceptually, person-centered and variable-

centered analyses look at two sides of the same coin. The statistical counterpart of this is that –

though factor and mixture models are not completely equivalent – “(…) a k-group mixture (or

latent profile) model can be redescribed as factor model with k-1 factors that is identical up to

first- and second-order moments” (Loken & Molenar, 2007, p. 293). Finally, as with other latent

variable models, the reification-fallacy should be avoided. Reification occurs when abstract

concepts represented by latent variables are taken too literally and are interpreted as if they had

a real, physical existence (Agresti, 2019; Bauer, 2007). Just because one successfully identified

latent classes and attached labels to them does not mean that these classes exist in the real world.
 31

In many cases, latent classes should be seen as statistical summaries that identify persons with

similar response patterns and, thus, can be helpful interpretative tools. However, such analytic

summaries are insufficient evidence to make the case that a specific typology has a concrete

reality or that an individual is of a certain naturally occurring type (Bauer, 2007; Masyn, 2013).

4.2 A note on software and further reading

There is a wealth of alternatives to the software introduced in this chapter. Many R packages

relevant for mixture modeling are listed on the Cran Task Views Psychometric Models and

Methods and Clusters (https://cran.r-project.org/web/views). As for proprietary software,

Mplus and Latent Gold are easy to use, can estimate a broad array of mixture models, and

implement modern approaches of handling covariates and distal outcomes. An applied

introduction to mixture modeling using Mplus is available, for example, in Wang and Wang

(2020). For Latent Gold, instructive user guides are freely available on the software’s website.

These and other software options are reviewed in Haughton et al. (2009), Hickendorf et al.

(2018), and Uebersax (2012).

For delving more deeply into mixture models, the following selected sources may be of

interest: Highly readable introductions can be found in Collins and Lanza (2010), Masyn

(2013), and Vermunt and Magidson (2002). Systematic, though more technically challenging,

overviews are available in Hancock and Samuelsen (2008), Hancock et al. (2019), McLachlan

and Peel (2000), Muthén (2008), and Sterba (2013). The edited volumes by Hagenaars and

McCutcheon (2002), and Rost and Langeheine (1997) contain high quality illustrative

applications (mostly of LCA) from a broad range of disciplinary and substantive perspectives.

References

Agresti, A. (2019). An introduction to categorical data analysis. Wiley.

Asparouhov, T., & Muthén, B. (2014). Auxiliary variables in mixture modeling: Three-step

approaches using Mplus. Structural Equation Modeling: A Multidisciplinary Journal,

21(3), 329-341. https://doi.org/10.1080/10705511.2014.915181

 32

Asparouhov, T., & Muthén, B. (2015). Residual associations in latent class and latent transition

analysis. Structural Equation Modeling: A Multidisciplinary Journal, 22(2), 169-177.

https://doi.org/10.1080/10705511.2014.935844

Asparouhov, T., & Muthén B. (2021). Auxiliary variables in mixture modeling: Using the BCH

method in Mplus to estimate a distal outcome model and an arbitrary second model

(Mplus Web Notes 21, Version 11). https://www.statmodel.com/MixtureModeling.shtml

Bakk, Z., & Kuha, J. (2021). Relating latent class membership to external variables: An

overview. British Journal of Mathematical and Statistical Psychology, 74(2), 340-362.

https://doi.org/https://doi.org/10.1111/bmsp.12227

Bauer, D. J. (2007). Observations on the use of growth mixture models in psychological

research. Multivariate Behavioral Research, 42(4), 757-786.

https://doi.org/10.1080/00273170701710338

Bauer, J., Gartmeier, M., Wiesbeck, A. B., Moeller, G. E., Karsten, G., Fischer, M. R., &

Prenzel, M. (2018). Differential learning gains in professional conversation training: A

latent profile analysis of competence acquisition in teacher-parent and physician-patient

communication. Learning and Individual Differences, 61(1), 1-10.

https://doi.org/https://doi.org/10.1016/j.lindif.2017.11.002

Bauer, J., & Mulder, H. R. (2013). Engagement in learning after errors at work: enabling

conditions and types of engagement. Journal of Education and Work, 26(1), 99-119.

https://doi.org/10.1080/13639080.2011.573776

Bauer, J. & Prenzel, M. (2021). For what educational goals do preservice teachers feel

responsible? On teachers’ ethos as professional values. In F. Oser, K. Heinrichs, J.

Bauer, & T. Lovat (Eds.), The international handbook of teacher ethos: Strengthening

teachers, supporting learners (pp. xxx). Springer. DOI:XXX

 33

Bergman, L. R., & Magnusson, D. (1997). A person-oriented approach in research on

developmental psychopathology. Development and Psychopathology, 9(2), 291-319.

https://doi.org/10.1017/S095457949700206X

Collins, L. M., & Lanza, S. T. (2010). Latent class and latent transition analysis. Wiley.

Finch, W. H., & Bronk, K. C. (2011). Conducting confirmatory latent class analysis using

Mplus. Structural Equation Modeling: A Multidisciplinary Journal, 18(1), 132-151.

https://doi.org/10.1080/10705511.2011.532732

Geiser, C. (2013). Data analysis with Mplus. Guilford.

Gillet, N., Morin, A. J. S., Mokounkolo, R., Réveillère, C., & Fouquereau, E. (2020). A person-

centered perspective on the factors associated with the work recovery process. Anxiety,

Stress, & Coping, 1-26. https://doi.org/10.1080/10615806.2020.1866174

Hagenaars, J. A., & McCutcheon, A. L. (Eds.). (2002). Applied latent class analysis. Cambridge

University Press.

Hancock, G. R., Harring, J. R., & Macready, G. B. (Eds.). (2019). Advances in latent class

analysis. IAP.

Hancock, G. R., & Samuelsen, K. M. (Eds.). (2008). Advances in latent variable mixture

models. IAP.

Haughton, D., Legrand, P., & Woolford, S. (2009). Review of three latent class cluster analysis

packages: Latent Gold, poLCA, and MCLUST. The American Statistician, 63(1), 81-

91. https://doi.org/10.1198/tast.2009.0016

Hickendorff, M., Edelsbrunner, P. A., McMullen, J., Schneider, M., & Trezise, K. (2018).

Informative tools for characterizing individual differences in learning: Latent class,

latent profile, and latent transition analysis. Learning and Individual Differences, 66, 4-

15. https://doi.org/https://doi.org/10.1016/j.lindif.2017.11.001

Hong, W., Bernacki, M. L., & Perera, H. N. (2020). A latent profile analysis of undergraduates’

achievement motivations and metacognitive behaviors, and their relations to

 34

achievement in science. Journal of Educational Psychology, 112(7), 1409–1430.

https://doi.org/10.1037/edu0000445

Jung, T., & Wickrama, K. A. S. (2008). An introduction to latent class growth analysis and

growth mixture modeling. Social and Personality Psychology Compass, 2(1), 302-317.

https://doi.org/https://doi.org/10.1111/j.1751-9004.2007.00054.x

Kunst, E. M., van Woerkom, M., & Poell, R. F. (2018). Teachers’ goal orientation profiles and

participation in professional development activities. Vocations and Learning, 11(1), 91-

111. https://doi.org/10.1007/s12186-017-9182-y

Linzer, D. A., & Lewis, J. B. (2011). poLCA: An R package for polytomous variable latent

class analysis. Journal of Statistical Software, 42(10), 1-29.

https://doi.org/10.18637/jss.v042.i10

Little, T. D., Slegers, D. W., & Card, N. A. (2006). A non-arbitrary method of identifying and

scaling latent variables in SEM and MACS models. Structural Equation Modeling: A

Multidisciplinary Journal, 13(1), 59-72. https://doi.org/10.1207/s15328007sem1301_3

Loken, E., & Molenaar, P. (2008). Categories or continua? The correspondence between

mixture models and factor models. In G. R. Hancock, & K. M. Samuelsen (Eds.),

Advances in latent variable mixture models (pp. 277–297). IAP.

Mair, P. (2018). Modern psychometrics with R. Springer. https://doi.org/10.1007/978-3-319-

93177-7

Masyn, K. (2013). Latent class analysis and finite mixture modeling. In T. D. Little (Ed.), The

Oxford handbook of quantitative methods in psychology (Vol. 2, pp. 551-611). Oxford

University Press.

McLachlan G. J. & Peel D. (2000). Finite mixture models. Wiley.

Muthén, B. (2008). Latent variable hybrids: Overview of old and new models. In Hancock, G.

R., & Samuelsen, K. M. (Eds.), Advances in latent variable mixture models (pp. 1-24).

IAP.
 35

Muthén, L. K., & Muthén, B. O. (1998-2017). Mplus user's guide (8th ed.). Muthén & Muthén.

Nylund, K. L., Asparouhov, T., & Muthén, B. O. (2007). Deciding on the number of classes in

latent class analysis and growth mixture modeling: a monte carlo simulation study.

Structural Equation Modeling, 14(4), 535–569.

https://doi.org/10.1080/10705510701575396

Nylund-Gibson, K., Grimm, R. P., & Masyn, K. E. (2019). Prediction from latent classes: A

demonstration of different approaches to include distal outcomes in mixture models.

Structural Equation Modeling: A Multidisciplinary Journal, 26(6), 967-985.

https://doi.org/10.1080/10705511.2019.1590146

Oberski, D. (2016). Mixture models: Latent profile and latent class analysis. In J. Robertson &

M. Kaptein (Eds.), Modern statistical methods for HCI (pp. 275-287). Springer.

https://doi.org/10.1007/978-3-319-26633-6_12

Pastor, D. A., Barron, K. E., Miller, B. J., & Davis, S. L. (2007). A latent profile analysis of

college students’ achievement goal orientation. Contemporary Educational Psychology,

32(1), 8-47. https://doi.org/https://doi.org/10.1016/j.cedpsych.2006.10.003

Peugh, J., & Fan, X. (2013). Modeling unobserved heterogeneity using latent profile analysis:

A monte carlo simulation. Structural Equation Modeling: A Multidisciplinary Journal,

20(4), 616-639. https://doi.org/10.1080/10705511.2013.824780

R Core Team (2021). R: A language and environment for statistical computing. R Foundation

for Statistical Computing, Vienna, Austria. https://www.R-project.org

Richter, D., Engelbert, M., Weirich, S., & Pant, H. A. (2013). Differentielle Teilnahme an

Lehrerfortbildungen und deren Zusammenhang mit professionsbezogenen Merkmalen

von Lehrkräften [Differential use of professional development programs and its

relationship to professional characteristics of teachers]. Zeitschrift für Pädagogische

Psychologie, 27(3), 193-207. https://doi.org/10.1024/1010-0652/a000104

 36

Rosenberg et al., (2018). tidyLPA: An R package to easily carry out latent profile analysis

(LPA) using open-source or commercial software. Journal of Open Source Software,

3(30), 978. https://doi.org/10.21105/joss.00978

Rost, J. (2003). Latent class analysis. In R. Fernandez-Ballesteros (Ed.), Encyclopedia of

psychological assessment (Vol. 1, pp. 539–543). Sage.

Rost, J. (2006). Latent-Class-Analyse. In F. Petermann & M. Eid (Eds.), Handbuch der

psychologischen Diagnostik (pp. 275–287). Hogrefe.

Rost, J., & Langeheine, R. (Eds.). (1997). Applications of latent trait and latent class models in

the social sciences. Waxmann.

Scrucca, L., Fop, M., Murphy, T. B., & Raftery, A. E. (2016, Aug). mclust 5: Clustering,

classification and density estimation using Gaussian finite mixture models. The R

Journal, 8(1), 289-317. https://doi.org/10.32614/RJ-2016-021

Scrucca, L., & Raftery, A. E. (2015). Improved initialisation of model-based clustering using

Gaussian hierarchical partitions. Advances in Data Analysis and Classification, 9(4), 447-

460. https://doi.org/10.1007/s11634-015-0220-z

Sterba, S. K. (2013). Understanding linkages among mixture models. Multivariate Behavioral

Research, 48(6), 775-815. https://doi.org/10.1080/00273171.2013.827564

Tofighi, D., & Enders, C. K. (2008). Identifying the correct number of classes in growth mixture

models. In G. R. Hancock & K. M. Samuelsen (Eds.), Advances in latent variable mixture

models (pp. 317–341). IAP.

Uebersax, J. (2012). LCA software. http://john-uebersax.com/stat/soft.htm

Vermunt, J. K., & Magidson, J. (2002). Latent class cluster analysis. In J. A. Hagenaars & A.

L. McCutcheon (Eds.), Applied latent class analysis (pp. 89–106). Cambridge University

Press.

Vermunt, J. K., & Magidson, J. (2016). Technical guide for Latent GOLD 5.1: Basic, Advanced,

and Syntax. Statistical Innovations Inc.

 37

Wang, J., & Wang, X. (2020). Structural equation modeling: Applications using Mplus (2nd

ed.). Wiley.